Email: [email protected]
Blender Chat: @ArenaGrenade
Github: https://github.com/ArenaGrenade
Website: https://arenagrenade.github.io/portfolio/ (Do not try it on low-end devices :), its still under construction so its not very optimal)
Twitter: https://twitter.com/agentAsskon
LinkedIn: https://www.linkedin.com/in/rohan-asokan/
For adding any feedback - please use the "Comment" button which would be one of the buttons on the top right corner. Thank you your time !!
This is an add-on to blender, that will see the introduction of machine learning based models and methods applied to fluid mechanics.
Machine learning models learn the behavior of certain simulation steps and its underlying principles. I would like to draw some parallels to how dynamic programming (Bottom Up or Tabulation form) works - we try to first fill in a table with the states and actions and then refer to them when we need them later and this method turns out to speed up a large section of problems. Machine learning models, learn this into a more compact representation, albeit losing out on a tiny fraction of the domain knowledge. Using this on novel fluid simulations that the neural network does not know about or has never seen - it tries to identify patterns and predicts using its knowledge base how the system will proceed. This means we don't have to preform the computationally intensive steps again - our machine learning model will be able to predict the behavior by identifying patterns.
Hence, we have a much faster method while losing out on the accuracy minimally. As is, numerical solutions are not 100% accurate and losing out on a tiny bit more of that accuracy for near real-time simulations, is the prime selling point of this idea.
This add-on would in fact be able to provide good quality fluid simulations on low-end PCs as well. As is clear from the papers I would be exploring as part of this project (given below) - the loss in accuracy is in most cases nonexistent, at least visually.
As speed is the main feature of this add-on, it allows for prototyping workflows where any user can mix and match the parameters and visualize them as fast as their thoughts go. This plugin can act as an intermediate stage between the idea for the simulation and production level simulations using other numerical methods.
By means of its sheer speeds, it would reduce the learning curve for creating great fluid effects and allow even beginners to get great fluid simulations by just trial and error.
To re-iterate: